Smaller cells have been the most important contributor to throughput improvement since the birth of cellular networks. They are likely to evolve further in the shift to cell-free massive MIMO (CF mMIMO), where multiple closely placed access points (APs) collaborate to serve users. This scheme is particularly suitable for millimeter wave (mmWave) communication, which enables very high data rates with its large bandwidth, but encounters severe challenges of high path loss and blockage. The CF mMIMO network is a good countermeasure to these two challenges by utilizing overlapping signals from different APs and macro-diversity. In this work, we demonstrate that mmWave CF mMIMO network optimization is largely an AP-user assignment problem. To solve this large-scale, nondifferentiable problem, we propose an unsupervised machine learning (ML) approach, which looks for the optimal solution autonomously without labels. A customized graph neural network architecture tailored to the problem properties is proposed, which enables distributed optimization without a central unit, allows for a varying number of users, and hierarchical permutation-equivariance of APs and users. A teacher-student model is applied to prune the graph, where the teacher model uses a fully connected graph for maximum performance, and the student model uses a pruned graph to reproduce the teacher's behavior with less communication in fronthaul. Moreover, a special training method is designed, which relaxes the combinatorial problem to a continuous one. In this way, we can apply gradient-based neural network training. An entropy-inspired penalty is introduced to make the relaxed problem equivalent to the original one. The analytical augmented Lagrangian method is combined with ML for the constrained optimization. Simulation results show that the proposed approach outperforms baselines in both performance and computation time. In addition, with a properly pruned graph, the proposed approach performs inference in a distributed manner with sparse message passing between APs, realizing a low signaling overhead in fronthaul, and a performance close to the fully connected graph.

Non-orthogonal multiple access (NOMA) is a promising multiple access technique. Its performance depends strongly on the wireless channel property, which can be enhanced by reconfigurable intelligent surfaces (RISs). In this paper, we jointly optimize base station (BS) precoding and RIS configuration with unsupervised machine learning (ML), which looks for the optimal solution autonomously. In particular, we propose a dedicated neural network (NN) architecture RISnet inspired by domain knowledge in communication. Compared to state-of-the-art, the proposed approach combines analytical optimal BS precoding and ML-enabled RIS, has a high scalability to control more than 1000 RIS elements, has a low requirement for channel state information (CSI) in input, and addresses the mutual coupling between RIS elements. Beyond the considered problem, this work is an early contribution to domain knowledge enabled ML, which exploit the domain expertise of communication systems to design better approaches than general ML methods.

space-division multiple access (SDMA) plays an important role in modern wireless communications. Its performance depends on the channel properties, which can be improved by reconfigurable intelligent surfaces (RISs). In this work, we jointly optimize SDMA precoding at the base station (BS) and RIS configuration. We tackle difficulties of mutual coupling between RIS elements, scalability to more than 1000 RIS elements, and high requirement for channel estimation. We first derive an RIS-assisted channel model considering mutual coupling, then propose an unsupervised machine learning (ML) approach to optimize the RIS with a dedicated neural network (NN) architectureRISnet, which has good scalability, desired permutation-invariance, and a low requirement for channel estimation. Moreover, we leverage existing high-performance analytical precoding scheme to propose a hybrid solution of ML-enabled RIS configuration and analytical precoding at BS. More generally, this work is an early contribution to combine ML technique and domain knowledge in communication for NN architecture design. Compared to generic ML, the problem-specific ML can achieve higher performance, lower complexity and permutation-invariance.

Millimeter wave (mmWave) cell-free massive MIMO (CF mMIMO) is a promising solution for future wireless communications. However, its optimization is non-trivial due to the challenging channel characteristics. We show that mmWave CF mMIMO optimization is largely an assignment problem between access points (APs) and users due to the high path loss of mmWave channels, the limited output power of the amplifier, and the almost orthogonal channels between users given a large number of AP antennas. The combinatorial nature of the assignment problem, the requirement for scalability, and the distributed implementation of CF mMIMO make this problem difficult. In this work, we propose an unsupervised machine learning (ML) enabled solution. In particular, a graph neural network (GNN) customized for scalability and distributed implementation is introduced. Moreover, the customized GNN architecture is hierarchically permutation-equivariant (HPE), i.e., if the APs or users of an AP are permuted, the output assignment is automatically permuted in the same way. To address this combinatorial problem, we relax it to a continuous problem, and introduce an information entropy-inspired penalty term. The training objective is then formulated using the augmented Lagrangian method (ALM). The test results show that the realized sum-rate outperforms that of the generalized serial dictatorship (GSD) algorithm and is very close to an upper bound in a small network scenario, while the upper bound is impossible to obtain in a large network scenario.

This work presents a machine learning approach to optimize the energy efficiency (EE) in an interference network. This optimization problem is non-convex and it is difficult to find its global optimum. We propose an unsupervised machine learning framework to approach the global optimum. While the training of the neural network (NN) takes moderate time, applying the trained model requires very low computational complexity. In particular, we introduce a novel objective function based on the reparameterization trick, which makes it possible to prune poor local optima to converge to the global optimum. Furthermore, we design a dedicated NN architecture SINRnet for signal-to-interference-noise ratio (SINR)-related optimization problems in interference networks, which is permutation-equivariant and classifies channels according to their positions in the SINR expression. In this way, we encode our domain knowledge into the NN design. Training and testing results show that the proposed method outperforms the successive convex approximation (SCA) algorithm, achieving an EE close to the global optimum found by the branch-and-bound algorithm and with reasonable computational effort. Thus, the proposed approach finds a balance between computational complexity and performance.

The reconfigurable intelligent surface (RIS) is a promising technology that enables wireless communication systems to achieve improved performance by intelligently manipulating wireless channels. In this paper, we consider the sum-rate maximization problem in a downlink multi-user multi-input-single-output (MISO) channel via space-division multiple access (SDMA). Two major challenges of this problem are the high dimensionality due to the large number of RIS elements and the difficulty to obtain the full channel state information (CSI), which is assumed known in many algorithms proposed in the literature. Instead, we propose a hybrid machine learning approach using the weighted minimum mean squared error (WMMSE) precoder at the base station (BS) and a dedicated neural network (NN) architecture, RISnet, for RIS configuration. The RISnet has a good scalability to optimize 1296 RIS elements and requires partial CSI of only 16 RIS elements as input. We show it achieves a high performance with low requirement for channel estimation for geometric channel models obtained with ray-tracing simulation. The unsupervised learning lets the RISnet find an optimized RIS configuration by itself. Numerical results show that a trained model configures the RIS with low computational effort, considerably outperforms the baselines, and can work with discrete phase shifts.

This work presents a machine learning approach to optimize the energy efficiency (EE) in a multi-cell wireless network. This optimization problem is non-convex and its global optimum is difficult to find. In the literature, either simple but suboptimal approaches or optimal methods with high complexity are proposed. In contrast, we propose an unsupervised machine learning framework to approach the global optimum. While the neural network (NN) training takes moderate time, application with the trained model requires very low computational complexity. In particular, we introduce a novel objective function based on stochastic actions to solve the non-convex optimization problem. Besides, we design a dedicated NN architecture SINRnet for the power allocation problems in the interference channel that is permutation-equivariant. We encode our domain knowledge into the NN design and shed light into the black box of machine learning. Training and testing results show that the proposed method without supervision and with reasonable computational effort achieves an EE close to the global optimum found by the branch-and-bound algorithm and outperform the successive convex approximation (SCA) algorithm. Hence, the proposed approach balances between computational complexity and performance.

With the increasing application of internet of things (IoT), the number of wirelessly transmitting devices is on a rise. It is important that the energy efficiency (EE) is maximized to reduce interference and save energy. This work explores the possibility of power control for maximum EE in wireless interference networks using reinforcement learning (RL) techniques. We apply the soft actor-critic (SAC) algorithm based on entropy regularization that allows to escape local optima and foster exploration. This enables us to solve the energy efficient power control problem with reduced complexity. We demonstrate that the obtained solutions are close to the global optimum. In contrast to supervised machine learning (ML) techniques, we do not need any kind of labeled data in the training phase. The model free approach and the unsupervised nature of RL therefore reduce the required computational effort and has a better scalability as a consequence.